# Autocorrelation in CO2 and Temperature Time Series

By Andy May

In my last post I plotted the NASA CO2 and the HadCRUT5 records from 1850 to 2020 and compared them. This was in response to a plot posted on twitter by Robert Rohde implying they correlated well. The two records appear to correlate because the resulting R2 is 0.87. The least square’s function used made the global temperature anomaly a function of the logarithm to the base 2 of the CO2 concentration (or ‘log2CO2‘). This means the temperature change was assumed to be linear with the doubling of the CO2 concentration, a common assumption. The least squares (or ‘LS’) methodology assumes there is no error in the measurements of the CO2 concentration and all error resulting from the correlation (the residuals) resides in the HadCRUT5 global average surface temperature estimates.

In the comments to the previous post, it became clear that some readers understood the computed R2 (often called the coefficient of determination), from LS, was artificially inflated because both X (log2CO2) and Y (HadCRUT5) were autocorrelated and increased with time. But a few did not understand this vital point. As most investors, engineers, and geoscientists know, two time series that are both autocorrelated and increase with time will almost always have an inflated R2. This is one type of “spurious correlation.” In other words, the high R2 does not necessarily mean the variables are related to one another. Autocorrelation is a big deal in time series analysis and in climate science, but too frequently ignored. To judge any correlation between CO2 and HadCRUT5 we must look for autocorrelation effects. The tool most often used is the Durbin-Watson statistic.

The Durbin-Watson statistic tests the null hypothesis that the residuals from a LS regression are not autocorrelated against the alternative that they are. The statistic is a number between 0 and 4, a value of 2 indicates non-autocorrelation and a value above 2 suggests negative autocorrelation, below 2 means positive autocorrelation. Since the computation of R2 assumes that each observation is independent of the others, we hope that we get a value of 2, that way the R2 is valid. If the regression residuals are autocorrelated and not random—that is normally distributed about the mean—the R2 is invalid and too high. In the statistical program R, this is done—using a linear fit—with only one statement, as shown below:

This R program reads in the HadCRUT5 anomalies and the log2CO2 values from 1850-2020 plotted in Figure 1, then loads the R library that contains the durbinWatsonTest function and runs the function. I only supply the function with one argument, the output from the R linear regression function lm. In this case we ask lm to compute a linear fit of HadCRUT5, as a function of log2CO2. The Durbin-Watson (DW) function reads the lm output and computes the DW statistic of 0.8 from the residuals of the linear fit by comparing them to themselves with a lag of one year.

The DW statistic is significantly less than 2 suggesting positive autocorrelation. The p value is zero, which means the null hypothesis that the HadCRUT5-log2CO2 linear fit residuals are not autocorrelated is false. That is, they are likely autocorrelated. R makes it easy to do the calculation, but it is unsatisfying since we don’t get much understanding from running it or from the output. So, let’s do the same calculation with Excel and go through all the gory details.

The Gory Details

The basic data used is shown in Figure 1, it is the same as Figure 2 in the previous post. Figure 1. The orange line is the log2CO2, use the right-hand scale. The multicolored line is the HadCRUT5 land + ocean global surface temperature record, it uses the left scale. The different colors identify the periods shown in the legend.

Strictly speaking, autocorrelation refers to how a time series correlates with itself with a time lag. Visually we can see that both curves in Figure 1 are autocorrelated, like most times series. What this means is that a large part of each value is determined by its preceding value. Thus, the log2CO2 value in 1980 is very dependent upon the value in 1979, and this is also true of the 1980 and 1979 values of HadCRUT5. This a critical point since all LS fits assume that the observations used are independent and that the residuals between the observations and the predicted values are random and normally distributed. R2 is not valid if the observations are not independent, a lack of independence will be visible in the regression residuals. Below is a table of autocorrelation coefficients for the curves in Figure 1 for time lags of one to eight years.

The autocorrelation values in Table 1 are computed with the Excel formula found here. The autocorrelation coefficients shown, like conventional correlation coefficients, vary from -1 (negative correlation) to +1 (positive correlation). As you can see in the Table both HadCRUT5 and log2CO2 are strongly positively autocorrelated, that is they are monotonically increasing, as we can confirm with a glance at Figure 1. The autocorrelation decreases with increasing lag, which is normally the case. All that means is that this year’s average temperature is more closely related to last year’s temperature than the year before and so on.

Row number one of Table 1 tells us that about 76% of each HadCRUT5 temperature and over 90% of each NASA CO2 concentration are dependent upon the previous year’s value. Thus, in both cases, each yearly value is not independent.

While the numbers given above apply to the individual curves in Figure 1, autocorrelation can clearly affect the regression statistics when the temperature and CO2 curves are regressed against one another. This bivariate autocorrelation is usually examined using the Durbin-Watson statistic mentioned above, and named for James Durbin and Geoffrey Watson.

Linear fit

As I did in the R program above, traditionally the Durbin-Watson calculation is performed on a linear regression of the two variables of interest. Figure 2 is like Figure 1, but we have fit LS lines to both HadCRUT5 and Log2CO2.

In Figure 2, orange is log2CO2 and blue is HadCRUT5. The residuals are shown in Figure 3, notice they are not random and appear to autocorrelate as we would expect from the statistics given in Table 1. They are autocorrelated and have the same shape, which is worrying.

The next step in the DW process is to derive a LS fit to the residuals shown in Figure 3, this is done in Figure 4.

Just as we feared, the residuals correlate and have a positive slope. Doing the DW calculations in this fashion, we get a DW statistic of 0.84, close to the value computed in R, but not exactly the same. I suspect that this is because the multiple sum-of-squares computations over 170 years of data leads to the subtle difference of 0.04. We can confirm this by performing the R calculation using the Excel residuals:

This confirms that both calculations match, but there were differences in the sum-of-squares calculations due to the different computer floating-point precision used in Excel and R. So, with a linear fit to both HadCRUT5 and log2CO2, there are serious autocorrelation problems. But both are concave upwards patterns, what if we used an LS fit that is more appropriate than a line? The plots look like a second order polynomial, let’s try that.

Polynomial Fit

Figure 5 shows the same data as in Figure 1, but we have fit 2nd order polynomials to each of the curves. The CO2 and HadCRUT5 data curve upward, so this is a big improvement over the linear fits above.

I should mention that I did not use the equations on the plot for the calculations, I did a separate fit to decades. The decades were calculated using 1850 as zero and 1850 to 1860 as decimal decades and so on to 2020 so that the X variable in the calculation had smaller values in the sum of squares calculations. This is to get around the Excel computer floating-point precision problem already mentioned.

The next step in the process is to subtract the predicted or trend value for each year from the actual value to create residuals. This is done for both curves, the residuals are plotted in Figure 6.

Figure 6 shows us that the residuals of the polynomial fits to HadCRUT5 and log2CO2 still have structure and the structure visually correlates, not a good sign. This is the portion of the correlation left, after the 2nd order fit is removed. In Figure 7 I fit a linear trend to the residuals. The R2 is less than in Figure 4

There is still a signal in the data. It is positive, suggesting that if the autocorrelation were truly removed with the 2nd order fit (we cannot say that statistically, but “what if”), there is still a small positive change in temperature, as CO2 increases. Remember, autocorrelation doesn’t say there is no correlation, it just invalidates the correlation statistics. If temperature is mostly dependent upon the previous year’s temperature, and we can successfully remove that influence, what remains is the real dependency of temperature on CO2. Unfortunately, we can never be sure we removed the autocorrelation and can only speculate that Figure 7 may be the true dependency between temperature and CO2.

The Durbin-Watson statistic

Now the calculations to compute the Durbin-Watson joint autocorrelation are done, but this time we used a 2nd order polynomial regression. Below is a table showing the Durbin-Watson statistic between HadCRUT5 and log2CO2 for a lag of one year. The calculations were done using the procedure described here.

The Durbin-Watson value of 0.9, for a one-year lag, confirms what we saw visually in Figures 5 and 6. The residuals are still autocorrelated, even after removing the second order trend. The remaining correlation is positive, as we would expect, presumably meaning that CO2 has some small influence on temperature. We can confirm this calculation in R:

Discussion

The R2 that results from a LS fit of CO2 concentration and global average temperatures is artificially inflated because both CO2 and temperature are autocorrelated time series that increase with time. Thus, in this case, R2 is an inappropriate statistic. R2 assumes that each observation is independent and we find that 76% of each year’s average global temperature is determined by the previous year’s temperature, leaving little to be influenced by CO2. Further, 90% of each year’s CO2 measurement is determined by the previous year’s value.

I concluded that the best function for removing the autocorrelation was a 2nd order polynomial, but even when this trend is removed, the residuals are still autocorrelated and the null hypothesis that they were not had to be rejected. It is disappointing that Robert Rohde, a PhD, would send out a plot of a correlation of CO2 and global average temperature implying that the correlation between them was meaningful without further explanation (as we showed in Figure 1 of the previous post) but he did.

Jamal Munshi did a similar analysis to ours in a paper in 2018 (Munshi, 2018). He notes that the consensus idea that increasing emissions of CO2 cause warming, and that the warming is linear with the doubling of CO2 (Log base 2) is a testable hypothesis. This hypothesis has not tested well because the uncertainty in the estimate of the warming due to CO2 (climate sensitivity) has remained stubbornly large for over forty years, basically ±50%. This has caused the consensus to try and move away from climate sensitivity toward comparisons of warming to aggregate carbon dioxide emissions, thinking they can get a narrower and more valid correlation with warming. Munshi continues:

“This state of affairs in climate sensitivity research is likely the result of insufficient statistical rigor in the research methodologies applied. This work demonstrates spurious proportionalities in time series data that may yield climate sensitivities that have no interpretation. … [Munshi’s] results imply that the large number of climate sensitivities reported in the literature are likely to be mostly spurious. … Sufficient statistical discipline is likely to settle the … climate sensitivity issue one way or the other, either to determine its hitherto elusive value or to demonstrate that the assumed relationships do not exist in the data.”

(Munshi, 2018)

While we used CO2 concentration in this post, many in the “consensus” are now using total fossil fuel emissions in their work, thinking that it is a more statistically valid quantity to compare to temperature. It isn’t, the problems remain, and in some ways are worse, as explained by Munshi in a separate paper (Munshi, 2018b). I agree with Munshi that statistical rigor is sorely lacking in the climate community. The community all too often use statistics to obscure their lack of data and statistical significance, rather than to inform.

The R code and Excel spreadsheet used to perform all the calculations in this post can be downloaded here.

Key words: Durbin-Watson, R, autocorrelation, spurious correlation

# Works Cited

Munshi, J. (2018). The Charney Sensitivity of Homicides to Atmospheric CO2: A Parody. SSRN. Retrieved from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3162520

Munshi, J. (2018b). From Equilibrium Climate Sensitivity to Carbon Climate Response. SSRN. Retrieved from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3142525

#autocorrelation, #co2, #durbin-watson, #temperature

## 18 thoughts on “Autocorrelation in CO2 and Temperature Time Series”

An excellent discussion of the modes of energy dissipation is to be found here: https://www.clepair.net/GreenhousClP-Eng.html

And for observation of the way things really are, that will appeal to all your friends, and provide a good laugh, this is hard to beat: https://www.youtube.com/watch?v=EPNhTVicEe4

What have you got in mind to adapt to the depreciation of the \$US, the GB pound, and perhaps the value of your super fund? Have you laid in a supply of candles, toilet paper and Spam?

1. Tinus,

I think that it is you who needs to improve your understanding of statistics. Andy has correctly described an issue involved in calculating correlation coefficient between two time series that are strongly autocorrelated. This is a well known problem in statistics – and unsurprisingly there are well established statistical methods for examining and quantifying this occurrence. Andy describes some of those.

The high correlation between temperature and CO2 concentration is shown by these statistical methodologies to be spurious. This does not mean that there is no relationship between the two, only that much more sophisticated methods are needed to extract the relationship.

Similar – and indeed even “better” – spurious correlations can be found. It seems that you have not read the Munshi 2018 paper cited above by Andy. In brief, that paper shows a strong correlation between CO2 concentration and annual homicides in England and Wales – the R2 value there is 0.89! Would you care to provide a causative explanation of how CO2 concentration affects the annual count of homicides in England and Wales?

Statistics is a tricky business and it takes time and effort to understand how to use it properly.

1. Autocorrelation analysis is aimed at finding periodicities (repeating p

Tinus, Stop citing wikipedia, it is not a reliable source. I will not go there. The science of “finding periodicities” is Fourier analysis, not autocorrelation, you seem to have confused the two. See here:
https://rundle.physics.ucdavis.edu/PHYGEO30/Fourier_Transforms.pdf

I have no idea what you are talking about in the rest of your comment. As I said earlier, I used statistics in my work for 42 years, I don’t see any evidence you understand it at all. Stay away from wikipedia, it is misleading you.

1. Again you are wrong.
Wikipedia also provides links to statistics handbooks. So you could check whether or not Wikipedia provides adequate explanations.

2. Your comment demonstrates a deep misunderstanding of autocorrelation and its consequences for statistical analysis. Autocorrelation is an extensively and well-understood phenomenon that very commonly applies to time series data. Statistics books and lecture courses deal with it extensively since time series analysis applies to many fields. There are a variety of measurements and techniques for revealing autocorrelation and for dealing with it.

It is interesting that you criticise both Andy and Munshi but you are never specific about where their errors lie. Simply badmouthing folk is unconvincing. To me, both Andy and Munshi use various standard techniques. It seems to me that you simply don’t like the results they produce, but they are what they are.

Munshi demonstrates that spurious correlations exist – it is a well known characteristic of autocorrelated time series and it must be guarded against carefully.

As for your harping on about the “simple physics mechanism” relating to CO2 – I keep repeating that the mechanism only applies in very limited circumstances to part of the Earth system. The Earth system is sufficiently complex that it could be the case that additional CO2 causes overall cooling of the surface, not heating. It almost certainly does cause cooling over Antartica – but it’s what happens in the 2/3rds of the Earth surface that is cloudy that is going to be crucial. The correlation between CO2 concentration and average Earth temperatures does not tell you much.

1. Tinus,

Well, at least your comment made me laugh. People can only debate matters if they first publish their views in a scientific journal – that is about the most ridiculous thing I’ve read in quite a while!

If your (mis)understanding is based on a “myriad” of publications in scientific journals, it is curious that you never seem to reference any of them. I’ve referred to a couple of papers already, one directly related to your “simple physics mechanism”, but you don’t seem to have noticed that.

I take the view that you’re hiding behind these “publications in scientific journals” as a way of covering up your lack of scientific or mathematical points in your postings.

As for understanding scientific publications procedures – I know those quite well, having published some papers in the past, including one (shared with many colleagues) in 1982 of which I am particularly proud. Subject matter, you ask? Physics, funnily enough.

2. Tinus, Mike is correct and I agree with him. I won’t repeat what he said. Just one comment:

Your ‘analysis’ of residues can tell you something on the less than 20% not covarying part of the measured CO₂ and temperature values, absolutely nothing on the covarying part. This analysis therefor is useless in your line of discussion.

The real discussion should be on the >80% covariance.

Residual analysis is not restricted to 20%, it can take an R^2 down to zero. I could put a full residual analysis into one post, there will be more on this topic later. Covariance is simply a measure of the correlation between two variables, but it has units and is not dimensionless like the correlation coefficient.

I don’t know about Mike but I used advanced statistics in my work for 42 years. Reading what you have written, I would judge your understanding of advanced statistics to be very poor. You don’t even seem to know the basics. Sorry, but that is what I see.

1. Noud Vermeulen says:

A far superior correlation exists between CO2 and adjustments. According to RealClimateScience (https://realclimatescience.com/corruption-of-the-us-temperature-record/) R^2= 0.977 for NOAA adjustments and CO2.
Admitted: not peer-reviewed, but a warning warning that it is a good idea to thoroughly inspect the quality of data, before starting a sophisticated scrutinity of autocorrelation between CO2 and temperature.
Moreover, the direction of causation is beyond doubt. I expect only minor differences between HADCRUT and NOAA.

1. Since both Andy and I are using long term globally averaged CO₂ concentrations, the “adjustments” of US temperature measurements are hardly, if at all, relevant. The world is much larger than the USA! These adjustments are very small as compared to the uncertainties in the data anyway.

1. Tinus, Your rhetoric is getting old and you are not contributing to the discussion, just being a troll. A challenge for you: Contribute something substantive, like Noud or Mike. A proper reference, not Wikipedia. Real proof or evidence you are correct or we are wrong.

Just saying we are wrong is empty, no one believes you. You have to put forward a substantive argument, do your homework!

I will delete all empty posts. You will know when you have produced a substantive one, because I will not delete it.

2. Tinus,
If you need some help understanding Autocorrelation and the techniques and methods relating to it, you could use this course from Penn State:

https://online.stat.psu.edu/stat510/

This course specifically deals with Time Series Analysis, which is the focus of concern here.

1. The link you provide very clearly shows that autocorrelation analysis is aimed at identifying periodicities (as I said earlier!). In the application of Andy it only identifies a long time constant (decennia).